DeepTutor: An Agent-Native Personalized Tutoring System from HKUDS

Key points

• Research target: HKUDS/DeepTutor, an open-source agentic personalized tutoring system from the HKU Data Science lab.
• Problem solved: General AI tutors forget prior sessions, lack learner profiles, and split tools across disconnected modes. DeepTutor addresses three break points: context fragmentation, memory loss, and tool/mode coupling.
• Core architecture: A five-layer personalization substrate (orchestrator, tools, capabilities, memory context, learner profile) that decouples tools from capabilities so any tool can serve any mode.
• Six learning modes sharing one thread: Chat, Deep Solve, Quiz, Deep Research, Math Animator, and Visualize operate on a shared conversational thread, shared knowledge-base references, and a shared learner profile.
• 3-stage Deep Solve pipeline: Personalized Investigation, Step-by-Step Solving (ReAct loop with adaptive replanning and context compression), and Evidence-Based Writing that adapts depth and tone to the learner's level.
• TutorBot: A long-running autonomous agent built on the lightweight nanobot runtime, with its own workspace, memory, skills, heartbeat/cron jobs, and session model. Multiple TutorBots can run in parallel and share one learner-profile substrate.
• Book Engine & Mimic quiz: Uploaded PDFs, Markdown, or text files are processed by OutlineAgent, RetrieveAgent, and CompileAgent into structured interactive pages with 13 content block types. Mimic mode parses past papers via MinerU and reproduces their style for exam preparation.
• Tech stack: FastAPI backend, React frontend, SQLite + Chroma local vector store, hybrid + naive RAG, multi-LLM support, Manim for math animation, Chart.js / Mermaid / SVG for visualization.
• Agent-native CLI: Commands such as deeptutor chat --kb=calculus --mode=deep_solve and deeptutor research --max_parallel=5 produce structured JSON for AI agents, with SKILL.md files letting other agents autonomously operate DeepTutor.
• Optional auth: Off by default; can be enabled with bcrypt password hashing, JWT sessions, an admin dashboard, and a PocketBase sidecar for OAuth and multi-user concurrency.
• Strengths: Unified context as a real UX innovation, deliberate 3-stage solve pipeline, agent-native evolution of TutorBot, exam-style Mimic quiz generation, and CLI/API exposure for other agents.
• Risks: Broad scope raises long-term maintenance burden; TutorBot adds concept overhead; the nanobot runtime lacks large-scale validation; local performance of multi-agent RAG and Manim rendering is not benchmarked.
• Comparison: Unlike Khanmigo (no persistent learner profile), Duolingo Max (language-specific), ChatGPT/Claude (no teaching loop), Anki (no AI tutoring), and Notion AI (writing assistant, no teaching loop), DeepTutor uniquely implements a full teach–practice–test–adapt loop as an agent-native open-source system.

References

• Repository: https://github.com/HKUDS/DeepTutor
• Paper: arXiv:2604.26962v1 — *DeepTutor: Towards Agentic Personalized Tutoring*
• Lab page: https://github.com/HKUDS

Core source files

• config/main.yaml — research settings
• config/agents.yaml — agent LLM parameters
• src/agents/question/ — question generation system
• src/agents/guide/ — guided learning system
• src/agents/co-writer/ — collaborative writing system
• deeptutor/tutorbot/skills/ — TutorBot skill definitions
• deeptutor/tutorbot/skills/skill-creator/SKILL.md — skill design philosophy